Break stalls when no bg work is happening #1884
Conversation
|
@IslamAbdelRahman has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
|
@IslamAbdelRahman updated the pull request - view changes - changes since last import |
|
@IslamAbdelRahman has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
db/db_impl.cc
Outdated
| } | ||
|
|
||
| delayed = true; | ||
| bg_cv_.TimedWait(static_cast<int>(stall_end - now)); |
There was a problem hiding this comment.
Can we still sleep a small amount of time? One reason I'm wary is to add one another use of TimedWait(). I recently realize that there may be corner case that a clock change may cause the wait not accurate. I think a shorter sleep will make it easier to read too.
There was a problem hiding this comment.
@siying, I think you mean there is a problem with NowMicros() not TimedWait(), correct ?
NowMicros() is the one that will be affected by system time change, correct ?
There was a problem hiding this comment.
@IslamAbdelRahman no I mean TimedWait(). I suspect inside pthread's wait until call, I think it calls non-monotonic time to determine duration to wait for.
db/db_impl.cc
Outdated
| // We will delay the write until the wall clock reach stall_end or | ||
| // we don't have any flushes or compactions running in the bg | ||
| uint64_t stall_end = sw.start_time() + delay; | ||
| while (bg_flush_scheduled_ > 0 || bg_compaction_scheduled_ > 0) { |
There was a problem hiding this comment.
What's the reason not callling write_controller_.NeedsDelay()?
|
Hey @siying, I think you missed my last comment here |
IslamAbdelRahman
force-pushed
the
new_stall
branch
from
e0c6b5c
to
44701ad
Compare
|
@IslamAbdelRahman updated the pull request - view changes - changes since last import |
|
Adress @siying comments
|
db/db_impl.cc
Outdated
| const uint64_t kDelayInterval = 10000; | ||
| uint64_t stall_end = stall_start + (delay * std::milli::den); | ||
| while (write_controller_.NeedsDelay()) { | ||
| if (env_->NowNanos() >= stall_end) { |
There was a problem hiding this comment.
I didn't mean we need to call a monotonic clock here. I just want to avoid TimedWait() if possible. But what you are doing here is still correct.
|
@IslamAbdelRahman has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
|
trying to figure out why the test is failing now |
IslamAbdelRahman
force-pushed
the
new_stall
branch
from
44701ad
to
bc89608
Compare
|
@IslamAbdelRahman updated the pull request - view changes - changes since last import |
|
@IslamAbdelRahman has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator. |
Current stall will keep sleeping even if there is no Flush/Compactions to wait for, I changed the logic to break the stall if we are not flushing or compacting
db_bench command used